How to set block sizes to maximize performance for matmul

[RuABraun]

Hi,
I'm interested in optimizing a matmul operations where I know the dims (e.g. 1024x512 @ 512x2048) and the GPU involved (A100).

I know I could just empirically test a lot of different options out, but I'd like to get a better of understanding of how CUDA and triton interact so that in other cases I have a better idea of what to do (I usually work in a context where some dims are variable).

Since I know the GPU, I know the number of SMs available and the size of the caches etc. might it be possible to get a guide on what values one should pick dependent on that?
Based on this I think I should be making sure that the BLOCK_SIZE_M and BLOCK_SIZE_N (to use the vars from the tutorial) result in tiles that optimize the use of the SMs. Is that going in the right direction?

I can find a lot of info about optimizing cuda kernels, but it's difficult to understand how that translates to writing triton code. Any help with that would be much appreciated!